Data mining for gearbox condition monitoring

Engineering datasets have growing rapidly in size and diversity as data acquisition technology has developed in recent years. However, the full use of the datasets for maximizing machine operation and design has not been investigated systematically because of the complexity of the datasets and huge amounts of data. This also means that data analysis based on traditional statistic based methods are no longer efficient in obtaining useful knowledge from these datasets. Thus this paper discusses dynamic and static datasets collected from a gearbox test rig with a typical drive system such that the datasets are considered representative for condition monitoring purposes. Dynamic datasets were analyzed to diagnose the condition of the gear: Healthy or Fault, using conventional signal processing techniques such as time-domain and frequency-domain analysis. The static data was also analyzed for comparative evaluation of detection performances. This procedure of data collection and analysis allowed a full understanding to be gained of condition monitoring datasets and paved the way for developing a more effective Data mining approach and efficient database. Moreover, to evaluate the effectiveness of using these new techniques, a prototype database was developed based on a gearbox test system and tested using these methods. The results obtained from a number of conventional methods have shown that data mining can obtain information for condition monitoring efficiently but not so accurately to give fault severity information, which is often sufficient for making maintenance decisions